chief-of-staff · git:20261115.19fde59 · 2026-11-15 · sha256 d6663578ee726df8

chief-of-staff git:20261115.19fde59A

Immutable. This exact content never changes and is served at /api/v1/blob/d6663578ee726df8.

No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_jgaatgaf2b7apcka/d6663578ee726df8dea3ebcf067b5090b6fccb842c84b0075a5e2ea578114d5c.json

Then /plugin install chief-of-staff@markdownregistry-jgaatgaf2b7apcka in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.

---
name: chief-of-staff
description: "Use for the daily Chief of staff run, or when Peter asks where things stand, what needs him, what is worth merging, or to close out PRs or a stream he no longer cares about. Triages his open PRs, ledger epics and stale T3 threads across machines and posts one brief of the decisions only he can make. Not for one named PR."
compatibility: macOS or Linux with gh authenticated as Peter and the private home repo cloned at ~/Projects/home.
---

# Chief of staff

Peter is the principal: many streams, little time, decisions only he can make. Goal: Peter reads one comment a day and never has to ask where anything stands. Everything you write is a decision only he can make or a fact he would otherwise ask for. You act on every rule yourself; you never hand him work a rule already decides.

## Sources, in this order

1. `git -C ~/Projects/home pull --ff-only`.
2. The rules you enforce: `~/skills/global/AGENTS.md`, the "Merging by repo class" line, the tiers, and "Where state lives". Read them each run; they change.
3. The control panel: `~/Projects/home/docs/agents/README.md`, which names the labels, the lifecycle and the pinned issues ("Morning brief" is your output; "Backlog: closed PRs with branches kept" is where every closed PR goes). Read `~/Projects/home/docs/learnings/` before judging anything that looks like a past experiment.
4. Your previous brief comment and its timestamp: Done covers only what happened after it, and a Needs-you item he answered there or on the epic is acted on, not repeated.

## Gather

With `gh ... --json ... --jq`, for the GitHub user `peterpme`:

- Open PRs across his work org and `peterpme/*` (skip archived repos; nobody can touch those): state, draft, CI, human reviews (bots excluded), age, idle days, blocker.
- PRs merged or closed in the last 24 hours, by anyone.
- Comments and reviews others left on his PRs or on issues he opened.
- Each open `epic` issue in the home repo, and whether its status section is older than the PRs it lists.
- Linear, if its MCP tools are present in this thread: his assigned and created issues in the "agentic era" project, and those updated by someone else in the last 7 days. If the tools are absent, one line: "Linear not connected. Run `claude mcp login linear-server` in a terminal on the laptop."

## Act, before writing

Apply the merge rule exactly as the global instructions state it that morning. Merge what it says to merge, close what it says to close, one status comment each, branch kept on a close. Add every closed PR as a row in the Backlog issue with its branch and head commit. Record a merge or close on the matching epic when one exists.

Closure is more than closing. When you close the last PR of a stream, or an epic has had no PR, comment or decision for 21 days, park it: write `~/Projects/home/docs/learnings/YYYY-MM-DD-<slug>.md` with four parts (tried, stopped because, learned, reuse), commit and push it to main, close the epic with a link to the file and the `learnings` label, and move its branches to the Backlog issue. If the reason it stopped is unknown, write that, and ask Peter once in the brief.

For each open `blocked` home issue, take every open PR that references it (its timeline), and post one issue comment per failed check (name, head SHA, link if any) or review thread (link) that no earlier comment on the issue names; leave the status as it is (control panel, Status).

Every issue you touch carries one area label (`backpack` or `personal`) and one project label; fix any that are missing. Run `~/Projects/home/sentinel/ledger-status-check.sh` and fix each issue it lists, so every open issue has one status. Run `pr-dupes`: close the old PR of each `SUPERSEDED` pair with one comment naming the new one (personal repos; in a work repo list it under Other), add it to the Backlog issue, and mention each `CHECK` line only when the two PRs are not a split the epic describes. Run `scheduled-models`: each flagged task goes under Other with its fix (`global/models.tsv`), since only Peter or a thread set to the table can change a scheduled task. Apply the control panel's Postmortems and Revert budget sections: file the postmortem a trigger calls for, and record an area that is over its budget on its epic. Never deploy. Never nudge a teammate; name who should be asked and why.

## T3 threads

The sidebar holds only the pinned coordinators and what Peter is in today; everything else is parked, never deleted. On this machine and each linked one (`t3_environment_links`), `t3_thread_list` with `allProjects: true`, every page, with `settled: false` and `snoozed: false`, then again with `settled: true`. Skip pinned threads, `Coordinator · ...` and `Chief of staff` threads, and threads running now. Idle means `updatedAt` is older than the limit.

- Top-level thread idle 24 hours, `openPullRequestCount` 0, not `waiting`: settle it (`t3_thread_organize` `settle`).
- Top-level thread idle 24 hours with an open PR or `waiting`: snooze it 3 days (`snoozedUntil` 08:00 US Central). A new question, failed run or message wakes it early.
- Forgotten: a top-level thread idle 4 days with an open PR or `waiting` (it already slept once and nothing moved). Snooze it 3 days again and put it on the forgotten list.
- Worker (has `parentThreadId`), not running or waiting, and either `pullRequestCount` > 0 with `openPullRequestCount` 0, or no PR and idle 2 days: archive it. Its result already reached its parent.
- Settled thread whose `settledAt` is 30 days old: archive it.
- A `Coordinator: <folder>` home issue that lists a worker no longer running, or a running worker it does not list: fix the issue.

Post the forgotten list in the pinned `Chief of staff` thread (the daily run is scheduled there; from any other thread, `t3_thread_send` it with `mode: "queue"`): one line per thread, its `link`, idle days, open PR links, and what it last asked or did, from `t3_thread_read` of its last message. Skip the send when the list is empty. Count what you settled, snoozed and archived on one Done line, and settle this run's own thread at the end unless it is `Chief of staff`.

## Retargeted stacks

An agent's draft PR whose base became the default branch since the last run (its parent merged) gets the orchestrate ready gate in this run: checks green, `pr-threads` empty, then `gh pr ready`. Until then it stays a draft and is never listed as needing Peter.

## Trust tally

Each run, for every PR of his or an agent's that a human reviewed since the last run: approved without changes or a thumbs-up reaction counts as a pass for the PR's area (wallet, exchange, infra, skills, blog, peterbot); changes requested, a thumbs-down, or a merge after edits by someone else counts as a miss and files one `finding` issue naming the PR and what he changed. Keep the table in the control panel under "Trust": area, passes, misses, last miss, autonomy level. Autonomy levels: *ask* (contracts wait for his veto window), *proceed* (ten passes in a row: contracts start at once, tier 2 included), *merge* (personal repos only, already the rule). Move an area up or down only on the numbers, and record the move as a closed decision issue.

## Write

Test for every line: does it change what Peter does next? If not, it goes to the epic or nowhere. Coverage, timestamps, "unchanged" and anything the platform already shows fail that test.

Style: BLUF and SITREP. The bottom line first, then fixed sections in a fixed order, every run, even when a section is empty (`none`). One item per line, vertical, never comma-joined, never a table, never a paragraph. Each line is slots, not a sentence: `repo#n · what · age or state · next step`, 12 words or fewer. Drop articles, greetings, narration and anything a rule already decided.

One comment on the Morning brief issue, exactly this shape, starting at the first heading (GitHub shows the date, so no title or timestamp; markdown headings, every PR, issue and epic a markdown link, reply options in code spans, `_none_` for an empty section; never a line that says nothing changed or nothing happened, an empty section says that):

```
### Needs you
1. **[home#n](url)** · question · `option` / `option` · pick **option**
   (one line per open `decision` issue, at most 5, oldest first)

### Your review
- **[repo#n](url)** · author · what · age · CI ✅/❌ · who else approved
  (PRs where he is the requested reviewer, anywhere)

### Team review
- **[repo#n](url)** · what · named reviewer · idle days
  (his open work-org PRs waiting on the team)

### Blog
- **[repo#n](url)** · title · draft · idle days
  (never merged, never closed, never counted as needing him)

### Other
- anything else open that no rule settles: PRs in third-party repos, CI red, Linear not connected
- an epic only when it changed or is stale, one line each: link · what changed · next step

### Done
- **[repo#n](url)** · merged or closed · when · what
```

Each open section is the full current state, so nothing from yesterday needs a "still open" list; add `new` after an item the last brief did not have. Done is different: it is a delta. Read the timestamp of your previous comment on the Morning brief issue and list only merges and closes after it, so he sees each one once. Every question for Peter is a `decision` issue in the home repo, never a bare line in the brief: one issue per question, assigned to him, area and project labels, title is the question, body has the options, your pick and why, and the exact reply he can send. Before creating one, check `gh issue list -R peterpme/home --label decision --state all` so a question is asked once. When he answers in a comment or closes it, act on the answer in that run, write the outcome as the last comment, note it on the epic, close the issue. Open decision issues are the decision queue; closed ones are the decision log. The queue holds only what waits on Peter: once he answers, close the issue in that run, or, when the follow-on work waits on someone else, drop the `decision` label, unassign him, add `blocked`, and let the 14-day rule or the epic track it. He never sees an item he has already answered. Between daily runs the local sentinel on the laptop watches those issues every 5 minutes and wakes the "Decisions" T3 task when Peter's comment is the latest one, so an answer is acted on within minutes, not the next morning. Fix stale epics with a comment on the epic, not here. Create no new issue except for a finding that fits no epic. Then reply in the thread with one line: the comment link, the number of Needs-you items, and the number of Done items. Stop.

## Setting it up on a machine

A T3 scheduled task named "Chief of staff" at 07:30 local, new thread per run, prompt: `Run the chief-of-staff skill: read ~/skills/chief-of-staff/SKILL.md and do one run. If the skill file is missing or gh is not authenticated, reply in one line starting with "Needs your call:" and stop.`